{
  "title": "Angel Marinov — Selected performance engineering studies",
  "author": {
    "name": "Angel Marinov",
    "handle": "0xamlab",
    "url": "https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#engineer",
    "github": "https://github.com/0xamlab",
    "linkedin": "https://www.linkedin.com/in/aimarinov/"
  },
  "portfolio_url": "https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/",
  "evidence_review_date": "2026-11-09",
  "page_updated_date": "2026-11-10",
  "scope": "Author-published workload-specific results. Benchmarks have not been independently rerun. Upstream statuses reflect the evidence snapshot, not a live status check.",
  "cases": [
    {
      "id": "cache",
      "category": "Inference",
      "eyebrow": "NVIDIA / SERVING",
      "title": "Persistent KV reuse",
      "value": "4.38",
      "unit": "×",
      "label": "batch wall-time speedup",
      "description": "Reuse a stored prefix in a fresh serving process instead of recomputing it.",
      "context": "A40 · Qwen2.5-7B · BF16 · 16K prefix · concurrency 8",
      "before": "23.66 s",
      "after": "5.40 s",
      "comparison": "versus forced prefix recomputation",
      "status": "Published benchmark",
      "date": "2026-11-08",
      "note": "Same-process native-hot cache was faster: 2.42 s. This is a persistent-cache result, not a blanket vLLM speedup. Three repetitions per cell; no statistical-significance claim.",
      "detail": "Fresh-process persistent-cache hits reduced mean batch wall time from 23.655112 s to 5.396244 s. Pooled TTFT p50 changed from 12.908762 s to 2.706828 s. The separate native-hot arm measured 2.418234 s mean wall time.",
      "role": "Built the external cache connector and its benchmark / lifecycle validation.",
      "sources": [
        {
          "label": "Benchmark record",
          "url": "https://github.com/thyn-ai/telys-vllm/blob/main/README.md"
        },
        {
          "label": "Detailed results",
          "url": "https://github.com/thyn-ai/telys-vllm/blob/main/bench/CONCURRENT.md"
        }
      ],
      "validation": "Published validation includes 25,414 / 25,414 16K blocks and 2,949 / 2,949 4K blocks. Bitwise restored-KV and 64-token output equality are scoped to batch-invariant mode and greedy decoding.",
      "quality": "Source reviewed; benchmark not independently rerun.",
      "portfolio_url": "https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-cache"
    },
    {
      "id": "gemm",
      "category": "Kernels",
      "eyebrow": "GPU / QUANTIZED MATMUL",
      "title": "Quantized matrix multiplication",
      "value": "38.1",
      "unit": "%",
      "label": "lower latency in the best tested cell",
      "description": "Change tile selection for large affine quantized matrix products.",
      "context": "M4 Max · 4-bit weights · BF16 activations · M=2048",
      "before": "28.483 ms",
      "after": "17.618 ms",
      "comparison": "versus the pinned upstream main branch",
      "status": "Open upstream PR",
      "date": "2026-11-03",
      "note": "Five tested configurations report 5.6–38.1% lower latency. Kernel-level measurements, not end-to-end model speedups. Submitted; not merged.",
      "detail": "N=12288, K=4096, group size 64. The patch adds 64×64 tiles for selected non-NAX, non-batched affine QMM paths while leaving other dispatch paths unchanged.",
      "role": "Authored the proposed tile-dispatch optimization and regression coverage.",
      "sources": [
        {
          "label": "Patch and measurements",
          "url": "https://github.com/ml-explore/mlx/pull/4826"
        }
      ],
      "validation": "The PR reports 982 tests passed / 58 skipped, plus 90 quantized-matmul subtests.",
      "quality": "Author-reported PR measurements; upstream acceptance not established.",
      "portfolio_url": "https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-gemm"
    },
    {
      "id": "fuzzy",
      "category": "Retrieval",
      "eyebrow": "MOJO / TEXT SEARCH",
      "title": "Fuzzy search at 100K documents",
      "value": "38.7",
      "unit": "×",
      "label": "warm-query speedup in the selected cell",
      "description": "Replace the JavaScript search hot path with compiled Mojo kernels.",
      "context": "M4 Max · 100K docs · 16-code-unit patterns · warm",
      "before": "1,239.508 ms",
      "after": "32.003 ms",
      "comparison": "versus Fuse.js 7.1.0",
      "status": "Published benchmark",
      "date": "2026-09-19",
      "note": "Warm results range from 8.8× to 38.7× across nine cells. The accelerated configuration uses 16 threads; this is not a single-thread comparison.",
      "detail": "Fixed-seed corpora with 15 patterns per cell. The source reports identical result order and zero score difference in the tested corpus. Cold index-build-plus-first-query measurements are separate.",
      "role": "Built the native acceleration path and API-compatible integration.",
      "sources": [
        {
          "label": "Benchmark record",
          "url": "https://github.com/thyn-ai/mojo-kernels/blob/main/README.md#fuse-mojo--fuzzy-search-fusejs-drop-in"
        }
      ],
      "validation": "The published corpus reports exactly zero score difference and identical result ordering. This does not assert support for every library option.",
      "quality": "Author-published synthetic-workload benchmark; not a universal search-speed claim.",
      "portfolio_url": "https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-fuzzy"
    },
    {
      "id": "bm25",
      "category": "Retrieval",
      "eyebrow": "MOJO / LEXICAL RANKING",
      "title": "BM25 versus an optimized scorer",
      "value": "1.10–1.95",
      "unit": "×",
      "label": "warm-query speedup across the tested grid",
      "description": "Precompute sparse scores, reduce allocation overhead, and batch queries.",
      "context": "M4 Max · 1K–100K docs · 5–20 query terms",
      "before": null,
      "after": null,
      "comparison": "versus the best measured bm25s / Numba variant",
      "status": "Published benchmark",
      "date": "2026-09-20",
      "note": "Scoring variants differ between libraries. Numerical parity is measured against rank_bm25, not bm25s. The smaller optimized-baseline range is the headline.",
      "detail": "Nine workload cells compare the best measured paths. The separate rank_bm25 comparison reports much larger 149×–13,137× ratios, but it is a different baseline and is not the headline here.",
      "role": "Built a compiled BM25 scorer with precomputed sparse scores and query batching.",
      "sources": [
        {
          "label": "Two-baseline comparison",
          "url": "https://github.com/thyn-ai/mojo-kernels/blob/main/README.md#benchmarks"
        }
      ],
      "validation": "Maximum reported absolute score difference versus rank_bm25: 3.6e-15. A limited top-10 overlap check with bm25s reports 1.00.",
      "quality": "Source-reported ratios are retained; rounded table cells should not be used to recalculate exact ratios.",
      "portfolio_url": "https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-bm25"
    },
    {
      "id": "scientific",
      "category": "Kernels",
      "eyebrow": "MOJO / SCIENTIFIC COMPUTE",
      "title": "Molecular wavefunction grids",
      "value": "72",
      "unit": "×",
      "label": "speedup versus the vectorized baseline",
      "description": "Evaluate a molecular-orbital grid with a compiled numerical kernel.",
      "context": "M4 Max · single thread · benzene / 6-31G* · 50³ grid",
      "before": "689 ms",
      "after": "9.5 ms",
      "comparison": "versus the cclib pyquante2 / NumPy path",
      "status": "Published benchmark",
      "date": "2026-09-19",
      "note": "A 50³ single-orbital grid, not an entire chemistry simulation. The much larger legacy-PyQuante ratio is intentionally not the headline.",
      "detail": "The selected comparison uses the NumPy-vectorized pyquante2 backend. Reported maximum absolute difference is 8.4e-13 for this workload.",
      "role": "Built the native Gaussian-grid evaluation path and Python integration.",
      "sources": [
        {
          "label": "Benchmark record",
          "url": "https://github.com/thyn-ai/mojo-kernels/blob/main/README.md#cclib-mojo--electron-density--wavefunction-grids-cclib-methodvolume"
        },
        {
          "label": "Baseline detail",
          "url": "https://github.com/thyn-ai/mojo-kernels/blob/main/python/cclib_mojo/README.md"
        }
      ],
      "validation": "Maximum reported absolute difference: 8.4e-13 against the selected vectorized baseline.",
      "quality": "Source rounds the speedup to 72×; absolute timings are rounded too.",
      "portfolio_url": "https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-scientific"
    },
    {
      "id": "vector",
      "category": "Retrieval",
      "eyebrow": "VECTOR SEARCH / TCP",
      "title": "Filtered vector retrieval",
      "value": "4,099",
      "unit": "QPS",
      "label": "at concurrency 8; recall reported as 1.0000",
      "description": "Use physical partitioning to avoid scanning unrelated vectors.",
      "context": "Cohere-100K · 768 dimensions · cosine · 1% filter",
      "before": null,
      "after": null,
      "comparison": "absolute throughput; no competitor-speedup claim",
      "status": "Open upstream PR",
      "date": "2026-07-21",
      "note": "Serial p99 is reported separately as 1.5 ms. It is not p99 at concurrency 8. Connector PR is open; this is not leaderboard certification.",
      "detail": "The author reports a real VectorDBBench CaseRunner against a live TCP service. The PR summary does not specify the benchmark hardware. Equality-label filtering is supported; numeric range filtering is not claimed.",
      "role": "Authored the benchmark connector and reported the filtered-search result.",
      "sources": [
        {
          "label": "Connector and reported results",
          "url": "https://github.com/zilliztech/VectorDBBench/pull/824"
        }
      ],
      "validation": "Reported recall 1.0000 and nDCG 1.0000 on the selected Cohere-100K workload.",
      "quality": "Author-reported result in an open PR; hardware is not specified in the inspected summary.",
      "portfolio_url": "https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-vector"
    },
    {
      "id": "h100",
      "category": "Inference",
      "eyebrow": "NVIDIA H100 / SERVING",
      "title": "27B-model concurrent serving",
      "value": "781.3",
      "unit": "tok/s",
      "label": "aggregate output throughput",
      "description": "Serve a 27B model at 50-way concurrency on one H100.",
      "context": "H100 80GB · Gemma 3 27B · FP16 · 550 in / 256 out",
      "before": null,
      "after": null,
      "comparison": "absolute serving result; no cross-engine multiplier",
      "status": "Archived result",
      "date": "2026-08-26",
      "note": "Taken from the supplied August 26, 2026 documentation snapshot. The live documentation could not be rechecked for this edition.",
      "detail": "The recorded open-loop throughput rises from 245.3 tok/s at concurrency 8, to 409.0 at 16, 585.6 at 32, and 781.3 at 50. Closed-loop batch numbers are not mixed into this result.",
      "role": "Built and benchmarked the serving runtime described in the source documentation.",
      "sources": [
        {
          "label": "Serving documentation",
          "url": "https://docs.algenta.ai/benchmarks/nvidia-llm-serving"
        }
      ],
      "validation": "Scope: 500-prompt sonnet-shaped workload, greedy decoding, prefix caching disabled, one H100 per engine.",
      "quality": "First-party archived document supplied by the user; live page and raw run were not revalidated.",
      "portfolio_url": "https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-h100"
    },
    {
      "id": "compute",
      "category": "Systems",
      "eyebrow": "RUNTIME / STRUCTURED COMPUTE",
      "title": "Warm deterministic execution",
      "value": "0.83–0.93",
      "unit": "ms",
      "label": "documented warm finance.npv execution",
      "description": "Keep a signed compute runtime resident between typed function calls.",
      "context": "One representative operation · warm resident runtime",
      "before": null,
      "after": null,
      "comparison": "documented latency observation, not a speedup",
      "status": "Documented observation",
      "date": null,
      "note": "Excludes first-ever runtime provisioning, documented at about 110 s. Not end-to-end agent latency or a benchmark of every function.",
      "detail": "The public documentation reports 0.83–0.93 ms for warm finance.npv. Hardware, sample count and a percentile distribution are not provided in the inspected passage.",
      "role": "Built the typed compute surface and managed-runtime integration.",
      "sources": [
        {
          "label": "Public documentation source",
          "url": "https://github.com/thyn-ai/sqai-mcp/blob/64e02f66768fbfb5f0da78e723e3932e86024d44/docs/faq.md"
        },
        {
          "label": "Technical docs",
          "url": "https://docs.sqai.com/faq"
        }
      ],
      "validation": "This entry is deliberately classified as an observation, not a matched benchmark campaign.",
      "quality": "Public documentation reviewed; hardware and distribution are unspecified.",
      "portfolio_url": "https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-compute"
    },
    {
      "id": "memory",
      "category": "Systems",
      "eyebrow": "RUST / MEMORY BEHAVIOR",
      "title": "Bounded phrase-query memory",
      "value": "19.51",
      "unit": "MB",
      "label": "peak RSS in the selected scored test",
      "description": "Replace explosive position expansion with a linear-memory matching path.",
      "context": "Repeated-token phrase stress case · scored search",
      "before": null,
      "after": null,
      "comparison": "prior implementation exceeded a 256 MiB watchdog",
      "status": "Open upstream PR",
      "date": "2026-11-06",
      "note": "The baseline is a watchdog lower bound, not a measured exact peak. No memory-reduction percentage is inferred. Submitted; not merged.",
      "detail": "For a repeated a/b/c stress document, the author reports 19.51 MB peak RSS in scored search after the prior implementation crossed a 256 MiB watchdog. A distinct adversarial case reports 19.68 MB.",
      "role": "Authored the proposed phrase-slop fix and memory-bounded matching path.",
      "sources": [
        {
          "label": "Patch and stress results",
          "url": "https://github.com/quickwit-oss/tantivy/pull/3248"
        }
      ],
      "validation": "The PR reports 163,840 exhaustive three-term comparisons against brute-force span matching and 1,353 library tests passed / 8 ignored.",
      "quality": "Author-reported stress measurement in an open PR; not a general database-memory result.",
      "portfolio_url": "https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-memory"
    }
  ],
  "recorded_serving_workloads": [
    {
      "label": "8K prefix · C=1",
      "prefix": "8K",
      "c": 1,
      "wall": [
        2.985708,
        2.91386,
        1.878551
      ],
      "ttft": [
        1.165793,
        1.104727,
        0.060614
      ],
      "tps": [
        21.435457,
        21.974601,
        34.06884
      ],
      "ratio": 1.0247
    },
    {
      "label": "8K prefix · C=4",
      "prefix": "8K",
      "c": 4,
      "wall": [
        6.572369,
        3.343736,
        2.029236
      ],
      "ttft": [
        3.96918,
        1.302441,
        0.117142
      ],
      "tps": [
        38.950953,
        76.573539,
        126.156782
      ],
      "ratio": 1.9656
    },
    {
      "label": "8K prefix · C=8",
      "prefix": "8K",
      "c": 8,
      "wall": [
        11.383825,
        3.716415,
        2.146576
      ],
      "ttft": [
        6.250244,
        1.548653,
        0.158475
      ],
      "tps": [
        44.976216,
        138.050123,
        238.52174
      ],
      "ratio": 3.0631
    },
    {
      "label": "16K prefix · C=8",
      "prefix": "16K",
      "c": 8,
      "wall": [
        23.655112,
        5.396244,
        2.418234
      ],
      "ttft": [
        12.908762,
        2.706828,
        0.252128
      ],
      "tps": [
        21.644457,
        94.97948,
        211.732989
      ],
      "ratio": 4.3836
    }
  ]
}